Ivo Alberink

Obtaining confidence intervals and Likelihood Ratios for body height estimations in images

In forensic practice, height estimations on perpetrators visible in video footage from surveillance cameras are regularly requested. There are several ways to do this. Insight is gained into the difference between actual and measured heights by taking validation measurements of a number of test persons. Variation between actual and measured heights is decomposed into a systematic part (because of height loss by pose, 3D modeling of the scene of crime, operator biases) and a random part (due to natural variation). On this basis a method is described for obtaining confidence intervals for the height, including head- and footwear, of questioned persons in images. Since the number of test persons is usually limited, the result is in terms of the Students t distribution. In addition, for cases in which a suspect is available, an expression is obtained for the Likelihood Ratio (LR), measuring the strength of evidence of resemblance of actual height of the suspect and measured height of the perpetrator. The Likelihood Ratio depends both on the rarity of the estimated perpetrators height and its closeness to the suspects height. Technical theorems included may be relevant for other forensic areas as well.

Fingermark evidence evaluation based on automated fingerprint identification system matching scores: the effect of different types of conditioning on likelihood ratios.

In recent studies, the evidential value of the similarity of minutiae configurations of fingermarks and fingerprints, for example expressed by automated fingerprint identification systems (AFIS), is determined by likelihood ratios (LRs). The paper explores whether there is an effect on LRs if conditioning takes place on specified fingers, fingerprints, or fingermarks under competing hypotheses: In addition, an approach is explored where conditioning is asymmetric. Comparisons between fingerprints and simulated fingermarks with eight minutiae are performed to produce similarity score distributions for each type of conditioning, given a fixed AFIS matching algorithm. Both similarity scores and LRs are significantly different if the conditioning changes. Given a common‐source scenario, “LRs” resulting from asymmetric conditioning are on average higher. The difference may reach a factor of 2000. As conditioning on a suspects finger(print) is labor‐intensive and requires a cooperating suspect, it is recommended to just condition on the number of minutiae in the fingermark.

Body Height Measurements in Images

Abstract:  For any measurement of a person’s height in an image, a difference exists between the actual height of the person and the image measurement. In order to gain knowledge about statistical behavior of differences between actual and measured heights it is necessary to make reference recordings, e.g., of test persons under the same recording conditions. To test whether the differences are dependent on camera and further circumstances, an experiment was set up which involved the measurement of 22 test persons using three cameras of varying quality. Reproducibility of measurements per image appears to be strongly dependent on the camera (quality), whereas systematic bias differs with the view point of the camera. Operator dependency of the measurement process is found, so its repetition by different operators is recommended.

Interoperator test for anatomical annotation of earprints.

ABSTRACT: As part of the Forensic Ear Identification (FearID) research project, which aims to obtain estimators for the strength of evidence of earmarks found on crime scenes, a large database of earprints (over 1200 donors) has been collected. Starting from a knowledge‐based approach where experts add anatomical annotations of minutiae and landmarks present in prints, comparison of pairs of prints is done using the method of Vector Template Matching (VTM). As the annotation process is subjective, a validation experiment was performed to study its stability. Comparing prints on the basis of VTM, it appears that there are interoperator effects, individual operators yielding significantly more consistent results when annotating prints than different operators. The operators being well trained and educated, the observed variation on both clicking frequency and choice of annotation points suggests that implementation of the above is not the best way to go about objectifying earprint comparison. Processes like the above are relevant for any forensic science dealing with identification (e.g., of glass, tool marks, fibers, faces, fingers, handwriting, speakers) where manual (nonautomated) processes play a role. In these cases, results may be operator dependent and the dependencies need to be studied.

Comparison of the Performance of Two Methods for Height Estimation

Abstract:  In the case study, two methods of performing body height measurements in images are compared based on projective geometry and 3D modeling of the crime scene. Accuracy and stability of height estimations are tested using reconstruction images of test persons of known height. Given unchanged camera settings, predictions of both methods are accurate. However, as the camera had been moved in the case, new vanishing points and camera matches had to be created for the reconstruction images. 3D modeling still yielded accurate and stable estimations. Projective geometry produced incorrect predictions for test persons and unstable intervals for questioned persons. The latter is probably caused by the straight lines in the field of view being hard to discern. With the quality of material presented, which is representative for our case practice, using vanishing points may thus yield unstable results. The results underline the importance of performing validation experiments in casework.

Repeatability and reproducibility of earprint acquisition.

Abstract:  For all forensic disciplines dealing with identification—e.g., of glass, tool marks, fibers, faces, fingers, handwriting, speakers—in which manual (subjective, nonautomated) processes play a role, operator dependencies are relevant. With respect to earprint identification, in the period 2002–2005, the Forensic Ear Identification research project collected a database of 1229 donors, three prints per ear, and laid down a “best practice” for print acquisition. Repeatability and reproducibility aspects of the print acquisition are tested. The study suggests that different operators may acquire prints of differing quality, with equal error rates of the matching system ranging from 9% to 19%. Moreover, it turns out that “matching” earprints are more alike when taken in a consecutive row than when taken on separate occasions. This underlines the importance of (1) studying operator effects, (2) operator training, and (3) not gathering “matching” reference material at the same occasion.

Posterior likelihood ratios for evaluation of forensic trace evidence given a two-level model on the data

In forensic science, in order to determine whether sets of traces are from the same source or not, it is widely advocated to evaluate evidential value of similarity of the traces by likelihood ratios (LRs). If traces are expressed by measurements following a two-level model with random effects and known variances, closed LR formulas are available given normality, or kernel density distributions, on the effects. For the known variances estimators are used though, which leads to uncertainty on the resulting LRs which is hard to quantify. The above is analyzed in an approach in which both effects and variances are random, following standard prior distributions on univariate data, leading to posterior LRs. For non-informative and conjugate priors, closed LR formulas are obtained that are interesting in structure and generalize a known result given fixed variance. A semi-conjugate prior on the model seems usable in many applications. It is described how to obtain credible intervals using Monte Carlo Markov Chain and regular simulation, and an example is described for comparison of XTC tablets based on MDMA content. In this way, uncertainty on LR estimation is expressed more clearly which makes the evidential value more transparent in a judicial context.

Variation in likelihood ratios for forensic evidence evaluation of XTC tablets comparison

In forensic comparison casework, where it is investigated whether items are from the same source, the likelihood ratio (LR) is a measure for evaluating the strength of evidence of the observed (dis)similarity between these items. The paper concentrates on evaluation of various LR models for comparison of XTC tablets from different production batches. Starting from a two‐level random effect model, where means are considered random, the distribution of these means may be assumed normal or approximated by kernel density estimations (KDEs), and variation in the data may or may not be allowed to differ per batch. It is investigated what effect the corresponding model and estimation choices have on the distribution of LRs for same‐ and different‐batch comparisons. Copyright

Measurement uncertainty when estimating the velocity of an allegedly speeding vehicle from images.

Abstract:  Sometimes the question arises whether it is possible to estimate the velocity of a speeding car recorded by closed circuit television cameras. By estimating the travelled distance of the car between two images and the time elapsed, estimation of the velocity is rather straightforward. However, to quantify the corresponding measurement uncertainty, the data analysis becomes more involved. The article describes two approaches as to how to derive the measurement uncertainty. In the first method, distance and timing are estimated separately, and the two uncertainties are combined to derive the measurement uncertainty for the velocity. For this, a frequentist and a Bayesian approach are described. In the second method, the measurement uncertainty for the speed is derived directly using validation recordings of a car driving by at known speed. The choice which method to use depends mainly on the length of the path that the car has travelled.

Performance Study of a Score-based Likelihood Ratio System for Forensic Fingermark Comparison

In this article, the performance of a score‐based likelihood ratio (LR) system for comparisons of fingerprints with fingermarks is studied. The system is based on an automated fingerprint identification system (AFIS) comparison algorithm and focuses on fingerprint comparisons where the fingermarks contain 6–11 minutiae. The hypotheses under consideration are evaluated at the level of the person, not the finger. The LRs are presented with bootstrap intervals indicating the sampling uncertainty involved. Several aspects of the performance are measured: leave‐one‐out cross‐validation is applied, and rates of misleading evidence are studied in two ways. A simulation study is performed to study the coverage of the bootstrap intervals. The results indicate that the evidential strength for same source comparisons that do not meet the Dutch twelve‐point standard may be substantial. The methods used can be generalized to measure the performance of score‐based LR systems in other fields of forensic science.